is liquidity related to prediction market efficiency? · correlated. several traders conduct...
TRANSCRIPT
Is Liquidity Related to Prediction Market Efficiency?
Paul C. Tetlock*
December 2007
Abstract
I investigate the relationship between liquidity and market efficiency using data from
short-horizon binary outcome securities listed on the TradeSports exchange. I find that
liquidity does not reduce—and sometimes increases—deviations of prices from financial
and sporting event outcomes. One explanation is that limit order traders are naïve about
other traders’ knowledge and unwittingly bet against them, which can slow the response
of prices to information. Consistent with this explanation, the limit orders that execute
during informative events have negative expected returns; and limit orders often execute
against traders who exploit the well-known favorite-longshot bias in prices.
* Please send all comments to [email protected]. I am currently visiting the finance department at
Yale, on leave from Texas. I am extremely grateful to TradeSports for their cooperation in allowing me to
collect the liquidity data, and their generosity in providing additional trading data. I thank Yakov Amihud,
Malcolm Baker, Ed Glaeser, Larry Glosten, John Griffin, Robert Hahn, Bing Han, Alok Kumar, Terry
Murray, Chris Parsons, Matt Spiegel, Laura Starks, Avanidhar Subrahmanyam, Sheridan Titman, Justin
Wolfers, Eric Zitzewitz, and seminar participants at Kansas and Texas for helpful comments.
1
I. Introduction
There is a growing body of theoretical and empirical research on securities
markets for contingent claims on uncertain events, often called prediction or information
markets. A key finding in this literature is that the prices in prediction markets can help to
produce forecasts of event outcomes with a lower mean squared prediction error than
conventional forecasting methods. For example, incorporating the prices in prediction
markets increases the accuracy of poll-based forecasts of election outcomes (Berg,
Forsythe, Nelson and Rietz (2003)), official corporate experts’ forecasts of printer sales
(Chen and Plott (2002)), and statistical weather forecasts used by the National Weather
Service (Roll (1984)).1
This study provides new evidence on how the forecasting accuracy of event
prediction markets depends on market liquidity. I use three years of intraday data on
one-day binary outcome securities based on sports and financial events traded on an
online exchange, TradeSports.com. I use three standard measures of securities liquidity,
low quoted bid-ask spreads, high market depth, and high trading volume. The idea is that,
in a liquid market, traders can cheaply buy and sell large quantities—e.g., O’Hara (1995).
I measure event forecasting accuracy as the difference between securities prices
and terminal cash flows. Unlike securities in conventional financial markets, I study
TradeSports’ securities that pay a single terminal cash flow at the end of their one-day
horizons, allowing me to repeatedly observe securities’ fundamentals (Thaler and Ziemba
(1988)). In addition, many of the securities are exposed to negligible systematic risks.
1 See Wolfers and Zitzewitz (2004) for a survey of the literature on prediction markets.
2
Thus, I can directly measure absolute pricing efficiency as the deviation of prices from
terminal cash flows.
I find that liquidity does not reduce—and sometimes increases—deviations of
securities prices from financial and sporting event outcomes. Prices can deviate from
outcomes because average prices differ from average event outcomes (i.e., poor
calibration), or because prices do not distinguish between events with different
probabilities (i.e., poor resolution). I find that the prices of liquid securities are not better
calibrated, and actually exhibit poorer resolution than the prices of illiquid securities.
These results are important for both direct and indirect reasons. Directly, the
findings address a key concern in using prediction markets prices to forecast events: their
accuracy could depend critically on market liquidity and trading activity. Several
researchers emphasize the potential of prediction markets to improve decisions (e.g.,
Hanson (2002), Hahn and Tetlock (2005), Sunstein (2006), and Cowgill, Wolfers, and
Zitzewitz (2007)). In principle, the range of applications is virtually limitless—from
helping businesses make better investment decisions to helping governments make better
fiscal and monetary policy decisions. For example, decision makers in the Department of
Defense, the health care industry, and multi-billion-dollar corporations, such as Eli Lilly,
General Electric, Google, France Telecom, Hewlett-Packard, IBM, Intel, Microsoft,
Siemens, and Yahoo, conduct internal prediction markets. The prices in these markets
reflect employees’ expectations about the likelihood of a homeland security threat, the
nationwide extent of a flu outbreak, the success of a new drug treatment, the sales
revenue from an existing product, the timing of a new product launch, or the quality of a
recently introduced software program.
3
Indirectly, the results shed light on the general relationship between liquidity and
absolute pricing efficiency, which could teach us something about financial markets.
Formally testing how liquidity and absolute pricing efficiency relate in conventional
financial markets would require that researchers observe securities terminal cash flows.
Even in cases where this is possible, such as in options or bond markets, other obstacles
exist. In options markets, inferences may depend critically on assumptions about the
market price of risk. In bond markets, there is often insufficient volatility in cash flows to
make the exercise interesting. By necessity, researchers in finance must rely on strong
assumptions relating absolute pricing efficiency to observable proxies for market
efficiency to estimate its relationship with liquidity. For example, many researchers argue
that liquidity increases market efficiency based on evidence from short-horizon return
predictability tests (e.g., Chordia, Roll, and Subrahmanyam (2006), and Chordia et al.
(2007)). Yet large deviations in prices from fundamentals could be associated with very
little or even no return predictability (Summers (1986)). Because nearly all efficiency
tests in equity markets do not employ measures of stocks’ fundamentals, one cannot infer
whether liquidity increases or decreases absolute pricing efficiency.
Despite these obstacles, assessing the relationship between liquidity and absolute
pricing efficiency remains important because absolute pricing efficiency—not return
predictability—determines whether capital is being efficiently allocated in financial
markets. A better understanding of the impact of liquidity on absolute pricing efficiency
could inform the decisions of firms and governments whose actions affect liquidity, and
thereby affect the efficient allocation of capital.
4
Beyond offering a particularly clean test of efficiency, there are other reasons to
expect that findings from TradeSports data are relevant for other financial markets. The
TradeSports exchange uses standard continuous double auctions, comparable to the
mechanisms used in the world’s major stock, currency, commodity and derivatives
exchanges.2 Many professional financial traders from New York, Chicago, and London
wager thousands of dollars in sports and financial markets on TradeSports.
Prices on TradeSports and conventional financial markets are likely to be
correlated. Several traders conduct automatic arbitrage transactions across conventional
exchanges and TradeSports, forcing similar pricing to prevail in these markets. Zitzewitz
(2006) reports that program trading accounts for over 95% of orders in the TradeSports
binary options securities on the daily Dow Jones index. Moreover, he finds surprising
evidence that some price discovery takes place in these daily Dow options. Specifically,
the implied volatilities from TradeSports binary options on the Dow can help forecast the
level and volatility of the Dow Jones index, even after controlling for multiple lags of
implied and historical volatilities from conventional CBOT, CBOE, and NYSE securities.
Arbitrage promotes the transmission of newly discovered price information from the
TradeSports exchange to other financial markets, and vice versa, within minutes.
There is also likely to be a common component in liquidity for similar contracts
traded on TradeSports and conventional financial exchanges because the popularity of
underlying securities indices is positively correlated across markets. For example, the
Dow securities are the most popular and actively traded on TradeSports; and the Dow
2 In both theory and practice, double auctions are particularly robust mechanisms that promote rapid
adjustment towards market equilibrium even in the presence of market frictions and trader irrationality—
e.g., Gode and Sunder (1993), Cason and Friedman (1996), Gjerstad and Dickhaut (1998), Noussair et al.
(1998), and Satterthwaite and Williams (2002).
5
stocks are also the most heavily traded stocks on the NYSE. Whatever motivates
investors to trade certain securities in one market may induce them to trade similar
securities in another market. Based on these similarities in liquidity and pricing across
markets, an understanding of the relationship between liquidity and efficiency on
TradeSports complements the indirect evidence from conventional financial markets,
where one generally cannot observe the terminal cash flows of highly volatile securities.
There are compelling theoretical reasons to think that liquidity and efficiency
could be either positive or negatively related. In many rational models of liquidity
provision, liquidity is beneficial to efficiency because it enhances the incentives for
traders to become informed and reveal their information through trading. Usually, the
indirect source of liquidity is random noise trading, which does not explicitly counteract
informed trading but does allow rational market makers to break even. An alternative
outside the scope of rational models is that some agents are naïve about the adverse
selection problem. If naïve agents submit near-market limit orders with insufficient
regard for the future release of information, they will unwittingly and systematically trade
with informed traders, effectively subsidizing liquidity provision. In such a market with
excessive liquidity provision, prices could respond more slowly to information.3
On TradeSports, I document two facts that are consistent with naïve liquidity
provision. Both facts could help to explain why forecasting accuracy is not higher in the
most liquid prediction markets, despite the strong incentives for information acquisition
and dissemination in liquid markets. First, limit orders that passively execute during
informative events have negative expected returns to expiration. This suggests that limit
3 Linnainmaa (2007) and Kaniel, Liu, Saar, and Titman (2007) report indirect evidence in financial
markets, but they cannot observe firms’ terminal cash flows. Baker and Stein (2004) and Tetlock and Hahn
(2007) present theoretical models in which liquidity subsidies lead to market underreaction.
6
orders retard the response of prices to new information and thereby inhibit forecasting
resolution. Second, limit orders traders passively buy low-priced securities—e.g., $4 and
below—and sell high-priced securities—e.g., $6 and above. Such limit orders could
sustain the overpricing of low-priced securities and the underpricing of high-priced
securities, a pattern known as the favorite-longshot bias (e.g., Thaler and Ziemba (1988)).
This second finding identifies a mechanism for how prices in liquid markets could remain
more poorly calibrated than prices in illiquid markets.
The layout of the paper is as follows. In Section II, I discuss the relationship
between this study and the prediction market and finance literatures. In Section III, I
describe the structure of the securities data from the TradeSports exchange and the
measures of liquidity used throughout the paper. In Section IV, I assess absolute pricing
efficiency in liquid and illiquid securities markets, focusing on forecasting calibration and
resolution. In Section V, I examine whether liquidity providers are naïve, documenting
the two key facts described above. In Section VI, I conclude and suggest directions for
further research on liquidity and securities market efficiency.
II. The Prediction Market Literature and Its Relationship to Finance
This study on prediction market efficiency is related to studies on the favorite-
longshot bias in wagering markets (e.g., Jullien and Selanie (2000), Wolfers and
Zitzewitz (2004), and Zitzewitz (2006)) and financial markets (e.g., Rubinstein (1985),
Brav and Heaton (1996), and Barberis and Huang (2007)). A key distinction is that these
prior studies do not link informational efficiency and market liquidity.
7
Several theoretical and empirical papers, however, do link these concepts. One
view is that illiquidity represents a transaction cost for informed arbitrageurs whose
trades make prices more efficient. For example, when liquidity increases in Kyle’s (1985)
model, informed traders bet more aggressively based on their existing information
because their trades have a smaller impact on prices. In addition, informed traders have
greater incentives to acquire more precise information in liquid markets. If informed
arbitrageurs are less active in illiquid markets where trading is expensive, securities’
prices in these markets may deviate by large amounts from their fundamental values. An
alternative view is that liquidity is a proxy for non-informational or noise trading, which
may harm informational efficiency.4 In behavioral finance models, various limits to
arbitrage prevent rational agents from making aggressive bets against noise traders—e.g.,
DeLong, Shleifer, Summers, and Waldmann (1990a) and Shleifer and Vishny (1997). If
liquid markets have more noise trading than illiquid markets and rational agents do not
fully offset noise traders’ systematic biases, then securities prices in liquid markets may
be inefficient relative to prices in illiquid markets.
Some recent papers provide indirect empirical support for the view that securities
mispricing is greater in illiquid markets—e.g., Wurgler and Zhuravskaya (2002),
Chordia, Roll, and Subrahmanyam (2006), and Chordia et al. (2007). Yet other papers
provide empirical evidence that suggests mispricing is greater in liquid markets—e.g.,
Bloomfield, O’Hara, and Saar (2007), Linnainmaa (2007), and Tetlock (2007). Of all
these papers, only Bloomfield, O’Hara, and Saar (2007) directly examines the deviation
of securities prices from fundamental values because the terminal cash flows of the
4 Variations in liquidity correspond to variations in noise trading in Kyle (1985), Glosten and Milgrom
(1985), and Baker and Stein (2004), but they may also result from variations in the search costs that buyers
and sellers incur in their efforts to transact—e.g., Duffie, Garleanu, and Pedersen (2005).
8
securities are unobservable in most real-world situations. The efficiency results from
TradeSports, where professional traders exchange large sums of money, demonstrate that
several laboratory results in Bloomfield, O’Hara, and Saar (2007) could generalize to
real-world financial markets.
III. Securities Data and Measures of Liquidity
I construct an automatic data retrieval program to collect comprehensive limit
order book and trading history statistics about each security traded on the TradeSports
exchange.5 The program runs at 30-minute intervals almost continuously from March 17,
2003 to October 23, 2006.6 All empirical tests in this paper include only data from the
single-day sports and financial securities that the program records. The vast majority of
TradeSports’ securities are based on one-day sports or financial events, such whether the
Yankees will win a particular baseball game or whether the Dow Jones Index will close
50 or more points above the previous day’s close. I focus on these securities to limit the
number of factors needed as controls in the statistical analysis that follows.7 Roughly
70% of TradeSports’ securities are based on sports events, 25% are based on financial
5 I am grateful to TradeSports Exchange Limited for granting me permission to run this program.
6 The program’s 30-minute interval is approximate because it records securities sequentially, implying that
the exact time interval depends on whether new securities have been added or subtracted and precise
download speeds. In practice, these factors rarely affect the time interval by more than two minutes. The
program stops running only for random author-specific events, such as software installations, operating
system updates, power failures, and office relocation—and technical TradeSports issues, such as daily
server maintenance and occasional changes in the web site’s HTML code. 7 For sports events, I consider only securities with an official TradeSports description that includes either
the text ―game,‖ ―bout,‖ or ―match‖; for financial events, I consider only securities based on the daily level
of stock indexes, which is the vast majority of all financial securities. Because almost all of the uncertainty
for these events is resolved during the scheduled event time, I focus on observations occurring within one
event duration of the event starting time—e.g., no more than 3 hours prior to the start of a 3-hour event.
9
events and fewer than 5% are based on events in all other categories combined—e.g.,
economic, political, entertainment, legal, weather, and miscellaneous.
The TradeSports exchange solely facilitates the trading of binary outcome
securities by its members, and does not conduct transactions for its own account.8
Securities owners receive $10 if a pre-specified, verifiable event occurs and $0
otherwise—e.g., the owner of the Dow Jones security mentioned above receives $10 if
and only if the index goes up by 50 points or more. For ease of interpretation, the
exchange divides its security prices into 100 points, worth $0.10 per point. The minimum
securities price increment, or tick size, ranges between 0.1 and one point in this sample.
TradeSports levies a commission of no more than 0.4% of the maximum
securities price ($10) on a per security basis whenever a security is bought or sold.9 At
the time of security expiration, when the payoff event is verifiable and the owner receives
payment, traders must liquidate all outstanding securities positions and incur
commissions. Note that the maximum $0.08 round-trip transaction fee is smaller than the
value of one point ($0.10) for most securities. This implies arbitrageurs have an incentive
to push prices back towards fundamental values if they stray by even one point.
Following conventions in other studies of financial markets, I eliminate
observations from TradeSports markets with poor quality price data. Specifically, I
exclude observations with a cumulative trading volume below 10 securities ($100),
market depth below 10 securities ($100), or bid-ask spreads exceeding 10% ($1.00).10
8 TradeSports limits the risk that the counterparty in a security transaction will default by imposing
symmetric margin requirements for each sale or purchase of a security by one of its members. Generally,
members must retain sufficient funds in their TradeSports account to guard against the maximum possible
loss on a transaction. TradeSports also settles and clears all transactions conducted on its exchange. 9 The exchange eliminated commissions for non-marketable limit orders on November 9, 2004.
10 The observations excluded by the combination of all three restrictions represent only 5% of volume on
the exchange. Using more stringent market activity criteria tends to strengthen the statistical results below.
10
These restrictions are designed to exclude securities without well-established market
prices for which tests of efficiency are not meaningful.
I use three measures of securities market liquidity: quoted bid-ask spreads, market
depth, and cumulative trading volume. I define the quoted spread as the difference
between the inside (lowest) ask and (highest) bid quotations. I compute market depth as
the sum of all buy and sell limit orders within the maximum 10% bid-ask spread, divided
by two. To enhance comparability across time periods and contracts, I use ad hoc cutoffs
of 1 point, 3 points, and 5 points—i.e., $0.10, $0.30, and $0.50—to partition observations
into groups sorted according to bid-ask spreads. The results are similar if I define
partitions based on the historical distribution of spreads during a rolling time window. I
follow an analogous procedure to sort securities by market depth and cumulative trading
volume, using ad hoc cutoffs of 100, 300, and 500 securities for both variables.
Table I shows that the three measures designed to capture liquidity exhibit some
similarity. The table reports the pair-wise correlations between the logarithms of quoted
spreads (LnSpread), market depth (LnDepth), and trading volume (LnVolm) for all
securities, sports securities and financial securities. All nine of the correlations in Table I
have the expected signs, and are statistically significant at the 1% level. For the group
including all securities, the volume correlations with spreads and depth are low because
the mean volume-based liquidity measure is higher in financial securities, whereas the
spread- and depth-based liquidity measures are higher in sports securities. The volume
correlations with spreads and depths are higher in the separate analyses of sports and
financial securities. The table also reports the average number of program observations
per security, which is between four and six for both sports and financial securities.
11
[Insert Table I here.]
IV. Tests of Market Efficiency
Now I analyze the absolute pricing efficiency of securities on the TradeSports
exchange. I conduct these tests separately for groups of securities sorted by each of the
three liquidity measures. I also analyze the efficiency of sports and financial markets
separately to investigate the possibility that pricing in these markets differs significantly.
Fortunately, the key results in this study apply to both sports securities and financial
securities, regardless of their exposure to market risk.11
A. Forecasting Calibration in Liquid and Illiquid Markets
Microstructure theory measures absolute pricing efficiency as the expected mean
squared error of prices minus cash flows, which can be decomposed into two components
as in Equation (1):
(1) E[(Payoff – Price)2] = E[Payoff – Price]
2 + E[(Payoff – 100 * Event Probability)
2]
First, I focus on measuring the squared bias component of absolute pricing efficiency
(i.e., the first term), as opposed to the conditional variance component (i.e., the second
term). This emphasis is standard practice in the literature on market efficiency in binary
prediction markets because expected cash flows are not directly observable for each
11
In unreported tests, I allow for the possibility that financial securities with positive exposure to market
risk have different expected returns from those with negative risk. I find a positive, but insignificant, risk
premium of less than 1% for the typical financial security with positive exposure to the market. This is not
surprising because three years of data is usually insufficient for estimating market risk premiums.
12
security. Thus, researchers cannot measure securities’ conditional variances without
making assumptions about biases, but they can estimate average biases by comparing
securities’ average cash flows to their prices. In this subsection, I estimate biases in prices
following conventions from related work—e.g., Wolfers and Zitzewitz (2004), Tetlock
(2004), and Zitzewitz (2006). In the next subsection, I make additional assumptions in an
effort to indirectly estimate the conditional variance component of mean squared error.
I employ a straightforward regression methodology to test the null hypothesis that
securities prices are unbiased predictors of securities’ cash flows. The null hypothesis is
that securities’ expected returns to expiration are zero, regardless of the current securities
price. The alternative hypothesis is that Kahneman and Tversky’s (1979) theory of
probability perception describes the pattern of expected returns across securities with
different current prices. A single observation consists of a security’s current price and its
returns until expiration. I measure current prices using the midpoints of the inside bid and
ask quotations to avoid the problem of bid-ask bounce that could affect transaction prices.
The results are robust to using the most recent transaction price instead.
I calculate a security’s percentage returns to expiration by subtracting its current
price from its payoff at expiration, which is either 0 or 100 points, then dividing by 100
points.12
This is the standard measure of returns in the prediction markets literature—e.g.,
Wolfers and Zitzewitz (2004), Tetlock (2004), and Zitzewitz (2006). Relative to
alternative measures that divide by price or the duration of the holding period, the
measure of returns to expiration described above possesses the advantages of being much
closer to homoskedastic and normally distributed, and it is symmetric for buyers and
12
I exclude the very small fraction of TradeSports contracts that do not expire at 0 or 100 points. I divide
by 100 points to represent the combined amount of capital that buyers and sellers invest in the security.
13
sellers. For the securities that I examine, the natural unit of information release is an
event, rather than a given amount of time. In addition, the opportunity cost of invested
funds is trivial over the daily time horizon of these securities.
The S-shaped form of the probability weighting function hypothesized in
Kahneman and Tversky (1979) and formalized in Prelec (1998) informs my choice of
pricing quantiles and statistical tests. The S-shape refers to a graph of subjective versus
objective probabilities, as shown first in Kahneman and Tversky (1979). Prelec (1998)
derives the S-shape in probability misperception from axiomatic foundations. His theory
predicts that agents overestimate the likelihood of events with objective probabilities less
than 1/e = 37% and underestimate the likelihood of events with objective probabilities
greater than 1/e. There is also an ample body of empirical evidence that is consistent with
a probability weighting function having a fixed point in the neighborhood of 1/e (Tversky
and Kahneman (1992), Camerer and Ho (1994), and Wu and Gonzalez (1996)).
Based on this evidence, I construct dummy variables, Price1 through Price5, for
five equally-spaced pricing intervals: (0,20), [20,40), [40,60), [60,80), and [80,100)
points. I then measure the returns until expiration for securities in each pricing quantile. I
test the null hypothesis that all returns to expiration are equal to zero against the
alternative that securities in the first two quantiles—Price1 and Price2—based on small
probability events (p < 40%), are overpriced and securities in the last three categories—
Price3, Price4, and Price5—based on large probability events (p ≥ 40%), are underpriced.
I report the results from three Wald (1943) tests based on this simple idea. The
first Wald test measures whether small probability events are overpriced on average:13
13
Despite the directional nature of over- and underpricing predicted by Kahneman and Tversky (1979) and
the favorite-longshot bias, I use two-tailed Wald tests to be conservative.
14
(2) (Price1 + Price2) / 2 = 0
The second Wald test assesses whether large probability events are underpriced:
(3) (Price3 + Price4 + Price5) / 3 = 0
The third Wald test measures whether large probability events are more underpriced than
small probability events—i.e., whether the mispricing function is S-shaped:
(4) (Price3 + Price4 + Price5) / 3 – (Price1 + Price2) / 2 = 0
Of the three, this is the most powerful test of the null hypothesis against the Kahneman
and Tversky (1979) alternative because it accounts for other factors that could influence
average returns in both small and large probability events.14
I also interpret equation (4)
as a test of the favorite-longshot bias: the expected returns on longshots, with prices less
than 40, are lower than the expected returns on favorites. In the original Ali (1977)
racetrack data, expected returns are roughly zero for horses with a 30% probability of
winning, and increase nearly monotonically with a horse’s probability of winning.
I use standard ordinary least squares to estimate the coefficients of the five pricing
categories. For all regression coefficients, I compute clustered standard errors, as in Froot
(1989), to allow for correlations in the error terms of all securities expiring on the same
calendar day, which simultaneously corrects for the repeated sampling of the same
security and the sampling of related events.15
This clustering procedure exploits the fact
that all event uncertainty is resolved on the day of expiration (see footnote 7).
To illustrate the efficiency tests and give an overview of the data, I first examine
the returns to expiration for all sports securities, all financial securities, and both groups
14
The choice of how to partition the pricing categories has little effect on the Wald tests because,
regardless of the partitioning, these tests assess whether the returns to expiration of securities priced below
40 points differ from the returns of securities priced above 40 points. 15
Using a finer clustering unit based on the expiration day and type of security does not affect the results.
15
together. Table II displays the regression coefficient estimates for Price1 through Price5
along with the three Wald tests described above. One interesting result is that neither
sports nor financial securities exhibit substantial mispricing, which is consistent with
Wolfers and Zitzewitz (2004) and Tetlock (2004).
[Insert Table II around here.]
The qualitative patterns in the pricing of both sets of securities and in their
aggregate suggest, however, that the favorite-longshot bias could play a role in any
mispricing that does exist. To aid the reader in identifying the S-shaped favorite-longshot
bias, Figure 1 provides a visual representation of mispricing in each pricing category for
sports, financial, and both types of securities.
[Insert Figure 1 around here.]
The securities based on small probability sports events in pricing quantiles 1 and
2 are slightly overpriced by 1.15 points (p-value = 0.127); and financial securities based
on large probability events in quantiles 3, 4, and 5 are underpriced by 2.21 points
(p-value = 0.030). The Wald test for the favorite-longshot bias rejects the null hypothesis
of no difference in expected returns at the 5% level for both sports and financial
securities. Interestingly, the magnitude of this bias decreases from an average of 2.13
points across the sports and financial groups to just 1.24 points in the aggregate group,
which is barely statistically significant at the 5% level. This reduction occurs because of
differences in the pricing patterns of sports and financial securities and the changing
relative composition of sports and financial securities within pricing quantiles.16
16
The disparity between the average of the individual estimates and the aggregate estimate illustrates the
importance of estimating the effects on sports and financial securities separately.
16
Having established that both sports and financial securities show a limited degree
of inefficiency on average, I now turn to the key test of whether the favorite-longshot bias
is more pronounced in illiquid or liquid securities. Panels A and B of Table III analyze
the favorite-longshot bias in sports and financial securities. The four regressions in each
panel in Table III sort securities according to their liquidity quantiles as measured by bid-
ask spreads. In both panels, the main result is that the favorite-longshot bias is no smaller
in liquid securities than the bias in illiquid securities. If anything, the bias appears larger
in liquid securities—e.g., the bias in liquidity quantiles 3 and 4 minus the bias in
quantiles 1 and 2 is positive (2.23%), but not statistically significant. Securities in the two
intermediate bid-ask spread quantiles exhibit a somewhat smaller favorite-longshot bias,
but this conclusion does not generalize for other liquidity measures. For the bid-ask
spread measure, the same patterns in the favorite-longshot bias apply to both sports and
financial securities.
[Insert Table III here.]
Based on the bid-ask spread measure, the financial securities are less liquid than
the sports securities. Panel B shows that relatively few financial securities have bid-ask
spreads less than $0.10, or 1% of the $10 maximum contract value. The financial
securities, however, are not less liquid based on the other liquidity measures, such as
limit order book depth and trading volume.
Next, I repeat the same regression analysis for sports and financial securities
sorted based on the other two liquidity measures: market depth and trading volume. For
brevity, Figure 2 summarizes only the favorite-longshot bias measures from these
17
analyses, together with the bid-ask spread results above. Figure 2 shows that the favorite-
longshot bias is virtually always positive, regardless of securities type or liquidity.
[Insert Figure 2 here.]
Overall, liquid and illiquid securities exhibit favorite-longshot biases of a similar
magnitude. In sports markets, liquid securities have slightly larger biases, but often not
significantly larger. In financial markets, the relationship between liquidity and the
favorite-longshot bias is quite weak, and depends on the liquidity measure. For example,
in financial securities with market depth in the top quantile, there is no favorite-longshot
bias; yet, in financial securities with bid-ask spreads in the bottom quantile, the favorite-
longshot bias exceeds 6%. This 6% bias is large relative to the bid-ask spreads of less
than 1% and the trading commissions of 0.8% or lower. More generally, the surprising
finding is that prices in liquid prediction markets are not better calibrated than prices in
illiquid markets.
A lack of statistical power does not seem to explain the lack of evidence for
liquidity improving efficiency. The favorite-longshot bias is consistently positive.
Moreover, the bias is actually larger in the top two liquidity quantiles than the bottom two
quantiles in five out of six cases in Figure 2. In addition, the confidence intervals are
sometimes narrow enough to reject the hypothesis that liquid securities are better
calibrated than illiquid securities at the 5% level—e.g., in the case of the trading volume
measure in sports securities. I offer a partial resolution to this puzzle in Section V.
In unreported tests, I evaluate the possibility that other security characteristics,
such as return volatility and time until expiration could explain the relationship between
mispricing and liquidity. The inclusion of the controls based on volatility and time
18
horizon does not significantly increase the explanatory power of the regressions in Table
III. Proponents of market efficiency can take comfort in the fact that few variables, other
than liquidity and price, are useful predictors of securities’ returns to expiration.
B. Forecasting Resolution in Liquid and Illiquid Markets
In this subsection, I measure how quickly securities prices converge toward event
outcomes and whether these rates depend on liquidity. In models of rational adverse
selection, periods of illiquidity tend to precede the release of information about
securities’ fundamental values and its incorporation into prices. Intuitively, traders are
reluctant to submit limit orders because they are wary that traders who submit market
orders possess information that they do not have. Thus, a lack of limit orders, or
illiquidity, could predict the release of information even if illiquidity does not cause
prices to incorporate information more rapidly.
In an effort to avoid this reverse causality problem, I estimate the impact of
liquidity on information release using instrumental variables (IV) for changes in liquidity.
Interestingly, the simple OLS estimates are qualitatively similar to the IV estimates,
suggesting that accounting for traders’ endogenous responses to time-varying adverse
selection does not affect the main findings. In Section V, I directly assess how limit order
traders respond to time-varying information release.
I use two instruments for market design that could have a significant impact on
prediction market liquidity, but do not have a direct impact on the release of event
information in sports and financial markets. On September 1, 2004, the TradeSports
19
exchange began a series of three changes in its trading fee structure. First, the exchange
reduced the trading fees on securities priced above $9.50 or below $0.50 by half. Within
two weeks, the exchange reduced the minimum tick size on its securities from $0.10 to
$0.01. Finally, on November 9, 2004, the exchange eliminated all trading fees on limit
orders that did not immediately execute. I construct two dummy variables to instrument
for changes in liquidity that occurred during the transition phase, September 1, 2004 to
November 9, 2004, and for changes occurring after all new rules were implemented. I use
observations on securities from March 1, 2004 through February 9, 2005 to evaluate the
impact of these changes.17
I measure the extent of information release during an event as the convergence of
prices toward securities’ terminal payoffs at expiration—i.e., the reduction in the squared
difference between prices and terminal payoffs. To analyze the change in this squared
difference, I use the identity in Equation (1) that decomposes mean squared error into
squared bias (calibration) and conditional variance (resolution) components.
The measurement of information release as the reduction in mean squared pricing
error (MSPE) requires an important assumption: reductions in MSPE occur only from
reduction in the conditional variance component—i.e., the squared bias component does
not change. This may be a reasonable approximation for short time horizons, but the
previous results suggest this assumption could fail for long time horizons because
squared bias approaches zero as an event ends. Specifically, if liquid markets exhibit a
greater favorite-longshot bias that is corrected over time, periods of liquidity will appear
17
Because an idiosyncratic author-specific event limits the number of observations in the pre-event period,
I use a six-month pre-event window to obtain a similar number of observations before and after the two
rule changes. The key regression coefficients are quantitatively sensitive to the specific event window
chosen, but the qualitative conclusions are robust.
20
to precede greater information release even if there is no causal relationship between the
two. This reasoning suggests my assumption biases the results toward finding that
liquidity causes prices to incorporate information more rapidly—but the bias is small
because there is not a large difference in mispricing between liquid and illiquid markets.
I use a two-stage least squares approach to examine how liquidity and the rate of
price convergence change as TradeSports changes its market design. I control for several
factors other than liquidity that influence the release of event information. For each sports
and financial security, I create a continuous scaled measure of the event time: event time
= current time – event start time / (event duration). I then create a series of dummy
variables to indicate the discrete event time position of a security in 20% increments,
where EvTime_X corresponds to an event time interval of ((X-1)*0.2, X*0.2] and X = 0, 1,
…, 6. Although some sports events last longer than expected, financial events almost
never do because they are based on index levels at pre-specified times. I also control for
the logarithm of the event time, in case the dummies fail to capture within-event trends in
liquidity or information release. In addition, I control for the logarithm of the date on
which the event is scheduled to expire to ensure that the excluded market design
instruments do not capture long-term trends in liquidity and information release.
Finally, I include a control for expected information release (ExpInfo) based on
the current securities price level. ExpInfo is defined as (100 – Price) * Price, which is the
expected value of squared pricing error under the assumption of market efficiency. To see
this, note that market efficiency implies prices equal true event probabilities, so that
Equation (5) holds:
21
(5) E[(Payoff – Price)2] = E[(Payoff – 100 * Event Probability)
2]
= (100 – Event Probability) * Event Probability = (100 – Price) * Price
In robustness checks, I find that the liquidity coefficient estimates are not particularly
sensitive to the use of alternative sets of control variables.
Table IV reports the results from six IV regressions each of which measures the
causal effect of liquidity on the reduction in MSPE over the next data recording period
(roughly 30 minutes). The six (2 x 3) IV regressions correspond to the results for sports
and financial securities for each of the three measures of liquidity: the logarithms of bid-
ask spreads (LnSpread), market depth (LnDepth), and trading volume (LnVolm). All
standard errors in Table IV are robust to heteroskedasticity and arbitrary intra-group
correlation within securities that expire on the same day (Froot (1989)).
[Insert Table IV here.]
Table IV consistently shows that an increase in liquidity slows the rate at which
prices incorporate event information. In four out of six IV regressions, I reject the null
hypothesis of no effect with a 1% Type I error; and, in the other two regressions, I reject
the null at the 10% level. The coefficient estimates are reasonably similar across the six
specifications. As bid-ask spreads decline, as market depth goes up, and as trading
volume increases, MSPE decreases more slowly. In all six specifications, the magnitude
of the effect is between 200 and 600 squared percentage points of MSPE per one standard
deviation increase in liquidity. To think about this magnitude, consider a security priced
at 55 that will eventually expire at 100, implying that the initial MSPE is 2025 points. A
drop in bid-ask spreads from 4 points to 2 points is roughly one standard deviation. On
average, this decrease in the spread would prevent the security priced at 55 from
22
converging to a price of 60—a new MSPE of 1600 points, which is 425 points lower—in
the next 30 minutes. In other words, the decrease in information release can be larger
than the increase in spreads.
Regression misspecification does not seem to explain this puzzling finding. Table
IV shows that I cannot reject any of the six overidentifying restrictions for the excluded
instrument at even the 10% level—see the Hansen (1982) J-statistics, which have a chi
squared distribution with one degree of freedom. In addition, the regression equations are
not underidentified: the Andersen (1984) likelihood-ratio statistic strongly rejects
underidentification in all six cases. Finally, a comparison of the Cragg-Donaldson
F-statistic to either the Cragg-Donaldson (1993) or the Stock and Yogo (2005) critical
values shows that the two instruments are not weak.
The intuition for the IV estimates is that the change in TradeSports’ market design
causes opposing changes in the speed of price responses and market liquidity. These
opposing changes occur in both sports markets and financial markets. Interestingly,
although TradeSports’ new rules increase all three measures of liquidity in financial
markets, they actually decrease all three measures of liquidity in sports markets (after
controlling for other factors). One possible explanation is that the optimal tick size differs
in sports and financial securities (e.g., Harris (1991) and Angel (1997)). As liquidity
increases in financial markets, prices respond more slowly to information; and, as
liquidity decreases in sports markets, prices respond more quickly to information.
23
V. Exploring the Liquidity Provision Mechanism
The evidence in Section IV suggests that forecasting calibration is not better in
liquid prediction markets, and that resolution is actually worse. This is surprising because
liquid markets generally provide greater incentives for traders to acquire and act on
information, whereas illiquid markets inhibit informed trading. In an effort to understand
the two main findings in Section IV, I study the mechanism whereby prices incorporate
information in prediction markets. I gather comprehensive trade-by-trade data from the
TradeSports Exchange Limited archive to monitor how prices respond to information.
This individual trade data indicates whether the buyer or seller initiated the trade.
Because the TradeSports exchange operates as a pure limit order book, I deduce that the
party initiating the trade submitted a market order (or a marketable limit order), and the
counterparty initially submitted a (non-marketable) limit order.
Two key properties of the limit orders that execute shed light on the liquidity
provision mechanism. First, during periods of frequent trading activity, limit orders that
execute have negative expected returns. This suggests that market orders during these
time periods convey valuable information that prices do not fully incorporate, perhaps
because limit orders naively stand in the way of price discovery. Second, limit order
traders passively buy low-priced securities and sell high-priced securities. This suggests
that traders submitting market orders exploit and counteract the favorite-longshot bias,
but that standing limit orders delay the correction in prices.
24
A. Limit and Market Order Behavior during Informative Events
First, I analyze the performance of active (market order) and passive (limit order)
trades when traders act on event-relevant information. Although I do not directly observe
event-relevant information, I develop a simple empirical proxy based on the clustering of
trades during time periods when the event is scheduled to be in-progress. The
TradeSports exchange supplies the official starting and ending event times—e.g., a
typical baseball game lasts 3 hours; and a typical financial index event lasts from 9:30am
Eastern time to 4:00pm Eastern time.
Table V summarizes the daily performance of market orders that trigger trades on
the TradeSports exchange. I define a time period to be informative if the preceding three
trades occurred within 60 seconds of each other and the event is currently in-progress. I
define a time period to be uninformative if the preceding three trades did not occur within
10 minutes of each other and the event is not yet in-progress. On each day, I aggregate
the properties of the trades in both groups from the perspective of the market order trader.
The value-weighted statistics weight trades by the number of securities traded, whereas
equal-weighted statistics weight trades equally. Table V reports the daily averages of
several trading statistics for the trades during informative periods, uninformative periods,
and all periods. I form these three groups separately within sports and financial securities.
The goal of this analysis is to assess whether traders submitting market orders or limit
orders receive and act on superior information during periods of high adverse selection.
[Insert Table V here.]
25
The main finding in Table V is that the trader submitting the market order
consistently outperforms the limit order trader following trade clusters during the event.
The magnitude of the market order trader’s value-weighted daily return is greater than
1% and highly statistically significant for both sports and financial securities. These
returns to expiration narrowly exceed the maximum round-trip commission costs of 0.8%
for market orders.
Interestingly, when the prior three trades are not clustered within a 10-minute
period and the event is not taking place, limit order traders perform significantly better
than market order traders. In these situations, the returns to expiration for limit order
traders are 0.77% and 1.12% for sports and financial securities, and are statistically
significant at the 5% level. In addition, limit order traders incur commissions of no more
than 0.4% upon expiration, and incur no commission on the initial trade after November
9, 2004. Thus, their returns to expiration are positive even after including fees.
The main result in Table V is robust to several alternative methodological
approaches. The second row of Table V shows that using equal-weighted returns, rather
than weighting each trade by the quantity traded, generates qualitatively similar results.
The third row (EW Return w/ Lag) shows that a market order trader can use public
information about past trades to exploit the market friction during informative time
periods. This row computes the hypothetical equal-weighted return for a trader taking the
same position in the current trade as the previous market order trader—e.g., taking the
buy-side of the current trade if the previous trade was buyer-initiated. The strongly
positive and significant return for this strategy shows that lags in price discovery prevent
complete adjustment to the direction of trade initiation. In unreported tests, I find that
26
prices only partially adjust to the initiation direction of the previous five trades. In
addition, sorting trades according to alternative proxies for information—e.g., using
recent price volatility instead of recent trading clustering—produces results qualitatively
similar to those in Table V.18
An implication of Table V is that limit order traders are systematically losing
money in time periods when prices are incorporating new information. It is as though
these traders pay insufficient attention to time-varying adverse selection. During
informative periods, limit order traders provide liquidity in excessive amounts relative to
a competitive market maker who tries to break even on each transaction. This naïve
liquidity provision, or failure to withdraw liquidity, delays the response of prices to
informative market order flows during the event. Still, during uninformative pre-event
periods, liquidity can be beneficial for market efficiency, as shown in columns Two and
Five in Table V: limit order traders prevent the overreaction of prices to uninformative
market orders during these periods. Figure 3 visually depicts these patterns in returns to
expiration around informative and uninformative events. The key point is that limit order
trades are associated with less efficient pricing during informative event periods, but
more efficient pricing during quiescent pre-event periods.
[Insert Figure 3 here.]
As an aside, the buy-sell imbalance row in Table V shows that there are more
buyer-initiated trades in both sports and financial securities, mainly during the pre-event
periods without trading clusters. I define buy-sell imbalance as the difference in the
quantities of buyer- and seller-initiated trades divided by the total quantity of trades.
18
Row four in Table V reveals that are far more market buy orders than market sell orders in pre-event
periods without clustered trades. This imbalance could indicate that aggressive pre-event traders suffer
from a framing bias, but I have no additional evidence to support this theory.
27
Because margin requirements and limits on trading are symmetric on the exchange, they
cannot explain the asymmetry between buyers’ and sellers’ aggressive market orders.
Thus, the positive buy-sell imbalance suggests that the preferences or prior beliefs of
aggressive pre-event traders are aligned with the way the TradeSports exchange frames
the events underlying securities—e.g., pre-event traders prefer to buy or speculatively
buy the security based on the Dow going up when the underlying event is framed as the
Dow going up, rather than down. However, the framing effect is much larger for
securities with negative risk exposure, suggesting that pre-event hedging explains much
of the effect in financial securities. Because TradeSports’ choices of event frames are not
random, it is unclear whether this framing effect is causal—but other researchers have
found a causal framing effect in similar contexts (Fox, Rogers, and Tversky (1996)).
B. Limit and Market Order Behavior across Pricing Quantiles
Next, I investigate whether limit orders could impede the response of prices to
aggressive market order traders who exploit market inefficiencies, such as the favorite-
longshot bias. The purpose of these tests is to reconcile the lack of evidence for better
forecasting calibration in liquid prediction markets in Section IV with the strong
incentives for informed trading that liquidity provides. Specifically, I explore the
possibility that limit order traders exhibit a systematic behavioral tendency that could
offset the beneficial impact of improvements in liquidity.
28
To test this hypothesis, I examine how the quantities of limit buy and sell orders
vary according to the prices of these binary-outcome (0 or 100 points) securities. I sort
both sports and financial securities into the five 20-point pricing quantiles described
earlier. For each of these ten (2x5) groups of securities, I aggregate the buy-sell
imbalance and returns over all trades taking place during scheduled event times. Again, I
aggregate trades on a daily basis from the perspective of the trader submitting the market
order. Panels A and B in Table VI report the value-weighted buy-sell imbalance and
returns for each of the pricing quantiles. I use the securities price from the previous
transaction as the basis for these sorts to avoid any mechanical correlation induced by
price pressure or bid-ask bounce. The results are slightly stronger if I use the
contemporaneous transaction price instead.
Table VI reveals that the patterns in the buy-sell imbalances of market orders are
similar to the favorite-longshot patterns in expected returns across pricing quantiles.
Market order traders in both sports and financial securities are more likely to sell
(longshot) securities with prices below 40 points, which have lower expected returns than
other securities. Market order traders tend to buy (favorite) securities priced above 60
points (or above 40 points in sports securities), which have higher expected returns than
other securities. The buy-sell imbalance results in Table VI demonstrate that limit order
traders may systematically, albeit passively, exacerbate the favorite-longshot bias in
prices. Executed limit orders tend to buy overpriced longshots and sell underpriced
favorites in both sports and financial securities. This tendency is strong in economic
terms: in sports securities purchases account for 55.6% and 43.7% of trades in pricing
29
quantiles 2 and 4, respectively; in financial securities, purchases account for 54.1% and
46.3% of trades in pricing quantiles 2 and 4, respectively.
Table VI also shows that market orders have significantly positive expected
returns in most pricing quantiles, suggesting market orders are based on superior
information. Expected returns in the middle pricing quantiles generally exceed expected
returns in the extreme quantiles. One interpretation is that the securities in the middle
quantiles have higher conditional variances—e.g., in an efficient market, conditional
variance is Price*(100 – Price), which is maximized at Price = 50 points—and limit
order traders do not set appropriately wide bid-ask spreads to account for this fact.
Lastly, I try to determine whether the favorite-longshot pattern in limit order
execution is active or passive by looking at order imbalances in the limit order book. Two
facts could explain the relatively large quantities of executed limit buy orders for
longshots: either limit traders submit more buy orders for longshots than for favorites, or
market order traders selectively sell to the limit order traders who submit buy orders for
longshots. In the first explanation, the buy-sell imbalance of unexecuted limit orders
should be higher for longshots than for favorites, whereas the second explanation predicts
a lower buy-sell imbalance for longshots. To test these explanations, on each day, I
compute the limit order book statistics using the same procedure described earlier for
trades. The only difference is that I use market depth rather than trading quantities to
generate value-weighted statistics.
[Insert Table VII here.]
Most of the evidence in Table VII supports the passive explanation of limit order
traders’ buy-sell imbalances, particularly in the financial securities shown in Panel B.
30
Sell-side limit orders are conspicuously missing in increasing amounts at higher pricing
quantiles, mimicking the patterns of buyer-initiated market order trades in Table VI. In
financial securities in Panel B of Table VII, the difference in buy-sell imbalances across
the pricing quantiles is highly statistically and economically significant—e.g., the
percentage of unexecuted limit sell orders is 12.4% lower in pricing quantile 1 than in
pricing quantile 5. The decreases in depth and increases in volume at higher pricing
quantiles reinforce the interpretation that buyer-initiated market orders actively take sell-
side liquidity away from high-priced securities. These trading patterns are less obvious in
the sports securities in Panel A of Table VII, though they appear within the top three
pricing quantiles. In the bottom two pricing quantiles, there is some evidence to suggest
that sports traders actively submit limit orders to purchase longshots.
[Insert Figure 4 here.]
Figure 4 summarizes the results on market and limit order buy-sell imbalances in
Table VI and Table VII. The solid and dashed black lines, describing financial market
order and limit order imbalances across the five pricing quantiles, correlate strongly with
each other (0.56). By contrast, the solid and dashed gray lines, describing sports market
order and limit order imbalances, are weakly positively correlated (0.13). In both cases,
the evidence is broadly consistent with the naïve and passive explanation of limit order
trader behavior.19
19
In unreported tests, I show that buy-sell imbalances in unexecuted limit orders do not predict event
outcomes, providing further confirmation that unexecuted limit orders are not informed.
31
VI. Concluding Discussion
This study’s main contribution is to illustrate the complex efficiency
consequences of liquidity provision, exploiting the simple structure of securities in
prediction markets with observable terminal cash flows and limited exposure to
systematic risks. In both sports and financial prediction markets, the calibration of prices
to event probabilities does not improve with increases in liquidity; and the forecasting
resolution of market prices actually decreases with increases in liquidity.
A careful examination of the behavior of limit order traders, who provide
liquidity, shows that these agents do not conform to the assumptions in standard rational
models. In times when market order flow is informative, limit order traders supply
liquidity in excessive amounts, systematically losing money to market order traders. This
finding could help to explain why prices in liquid markets are slow to incorporate event
information. Furthermore, limit order traders passively trade against the favorite-longshot
bias, possibly explaining why prices in liquid prediction markets are not better calibrated
than prices in illiquid markets.
More broadly, these results show that liquidity providers are not always rational,
and do not always attain zero profits. The findings also suggest that traders’ passivity and
naïveté partially explain underreaction to information in prediction markets. Given that
Linnainmaa (2007) observes similar behavioral and pricing phenomena in financial
markets, the same limit order mechanism could hinder efficiency in financial markets
with unusually high depth.
32
Anecdotally, cases such as technology stocks during the Internet boom raise
doubts about whether pricing in liquid markets is always more efficient than pricing in
illiquid markets. Further exploration of the limit order mechanism identified here and in
Linnainmaa (2007) could yield insights into these puzzling cases. The effect of excessive
liquidity provision on equilibrium prices depends critically on how arbitrageurs respond
to liquidity. Because liquidity gives arbitrageurs the option to enter and exit their trading
positions easily, it could induce them to take short-run outlooks and conduct trades that
actually exacerbate mispricing (e.g., DeLong et al. (1990b) and Brunnermeier and Nagel
(2004)). In other cases in which arbitrageurs have incentives to correct mispricing, they
may be unable to do so because they have less capital than naïve liquidity providers.
Theoretical models of the limit order mechanism could make new predictions about the
situations in which liquid markets are more or less efficient than illiquid markets.
33
References
Ali, Mukhtar M., 1977, Probability and utility estimates for racetrack bettors, Journal of
Political Economy 85, 803-815.
Anderson, T.W., 1984. Introduction to Multivariate Statistical Analysis. 2nd
ed. New
York: John Wiley & Sons.
Angel, James J., 1997, Tick size, share price, and stock splits, Journal of Finance 52,
655-681.
Baker, Malcolm, and Jeremy Stein, 2004, Market liquidity as a sentiment indicator,
Journal of Financial Markets 7, 271-299.
Barberis, Nicholas, and Ming Huang, 2007, Stocks as lotteries: the implications of
probability weighting for security prices, NBER Working Paper #12936.
Berg, Joyce, Robert Forsythe, Forrest Nelson and Thomas Rietz, 2003, Results from a
Dozen Years of Election Futures Markets Research, University of Iowa working
paper.
Bloomfield, Robert, Maureen O’Hara, and Gideon Saar, 2007, How noise trading affects
markets: an experimental analysis, Cornell University Working Paper.
Brav, Alon, and John B. Heaton, 1996, Explaining the underperformance of initial public
offerings: a cumulative prospect utility approach, Duke University Working
Paper.
Brunnermeier, Markus K., and Stefan Nagel, 2004, Hedge funds and the technology
bubble, Journal of Finance 59, 2013-2040.
Camerer, Colin and Teck H. Ho, 1994, Violations of the betweenness axiom and
nonlinearity in probability, Journal of Risk and Uncertainty 8, 167-196.
Cason, Timothy and Daniel Friedman, 1996, Price formation in double auction markets,
Journal of Economic Dynamics and Control 20, 1307-1337.
Chen, Kay-Yut and Charles Plott, 2002, Information Aggregation Mechanisms:
Concepts, Design, and Implementation for a Sales Forecasting Problem, CalTech
Working Paper No. 1131.
Chordia, Tarun, Richard Roll, and Avanidhar Subrahmanyam, 2006, Liquidity and
market efficiency, Emory University Working Paper.
34
Chordia, Tarun, Amit Goyal, Gil Sadka, Ronnie Sadka, and Lakshmanan Shivakumar,
2007, Liquidity and the post-earnings-announcement drift, Emory University
Working Paper.
Cowgill, Bo, Justin Wolfers, and Eric Zitzewitz, 2007, Prediction Markets Inside the
Firm: Evidence from Google, University of Pennsylvania Working Paper.
Cragg, J.G., and S.G. Donaldson, 1993, Testing identfiability and specification in
instrumental variables models, Econometric Theory 9, 222-240.
DeLong, J. Bradford, Andrei Shleifer, Lawrence H. Summers, and Robert J. Waldmann,
1990a, Noise trader risk in financial markets, Journal of Political Economy 98,
703-738.
DeLong, J. Bradford, Andrei Shleifer, Lawrence H. Summers, and Robert J. Waldmann,
1990b, Positive feedback investment strategies and destabilizing rational
speculation, Journal of Finance 45, 379-395.
Duffie, Darrell, Garleanu, Nicolae, and Lasse Heje Pedersen, 2005, Over-the-counter
markets, Econometrica 73, 1815-1847.
Fox, Craig R., Brett A. Rogers, and Amos Tversky, 1996, Options traders exhibit
subadditive decision weights, Journal of Risk and Uncertainty 13, 5-17.
Froot, Kenneth A., 1989, Consistent covariance matrix estimation with cross-sectional
dependence and heteroskedasticity in financial data, Journal of Financial and
Quantitative Analysis 24, 333-355.
Gjerstad, Steven, and John Dickhaut, 1998, Price formation in double auctions, Games
and Economic Behavior 22, 1-29.
Glosten, Lawrence, and Paul Milgrom, 1985, Bid, ask and transaction prices in a
specialist market with heterogeneously informed traders, Journal of Financial
Economics 14, 71-100.
Gode, Dhananjay K., and Shyam Sunder, 1993, Allocative efficiency of markets with
zero-intelligence traders: markets as a partial substitute for individual rationality,
Journal of Political Economy 101, 119-137.
Hahn, Robert W., and Paul C. Tetlock, 2005, Using information markets to improve
public decision making, Harvard Journal of Law and Public Policy 29, 214-289.
Hansen, Lars, 1982, Large sample properties of the generalized method of moments
estimators, Econometrica 50, 1029-1054.
35
Hanson, Robin, 2002, Decision markets, in Entrepreneurial Economics: Bright Ideas
from the Dismal Science, Oxford University Press.
Harris, Lawrence, 1991, Stock price clustering and discreteness, Review of Financial
Studies 4, 389-415.
Jullien, Bruno, and Bernard Selanie, 2000, Estimating preferences under risk: the case of
racetrack bettors, Journal of Political Economy 108, 503-530.
Kahneman, Daniel, and Amos Tversky, 1979, Prospect theory: an analysis of decision
under risk, Econometrica 47, 263-291.
Kaniel, Ron, Shuming Liu, Gideon Saar, and Sheridan Titman, 2007, Individual investor
trading around earnings announcements, Duke University working paper.
Kyle, Albert S., 1985, Continuous auctions and insider trading, Econometrica 53, 1315-
1336.
Linnainmaa, Juhani, 2007, The limit order effect, University of Chicago working paper.
Newey, Whitney K., and Kenneth D. West, 1987, A simple, positive-definite,
heteroskedasticity and autocorrelation consistent covariance matrix,
Econometrica 55, 703-708.
Noussair, Charles, Stephane Robin, and Bernard Ruffieux, 1998, The effect of
transactions costs on double auction markets, Journal of Economic Behavior and
Organization 36, 221-233.
O’Hara, Maureen, 1998, Market Microstructure Theory, John Wiley & Sons, Inc.
Prelec, Drazen, 1998, The probability weighting function, Econometrica 66, 497-527.
Roll, Richard, 1984, Orange juice and weather, American Economic Review 74, 861-880.
Rubinstein, Mark, 1985, Nonparametric tests of alternative option pricing models using
all reported trades and quotes on the 30 most active CBOE option classes from
August 23, 1976 through August 31, 1978, Journal of Finance 40, 455-480.
Sadka, Ronnie, and Anna Scherbina, 2007, Analyst disagreement, mispricing, and
liquidity, Journal of Finance 62, 2367-2403.
Satterthwaite, Mark A., and Steven R. Williams, 2002, The optimality of a simple market
mechanism, Econometrica 70, 1841-1863.
Shleifer, Andrei, and Robert W. Vishy, 1997, The limits of arbitrage, Journal of Finance
52, 35-55.
36
Stock, James H., and Motohiro Yogo, 2005, Testing for weak instruments in linear IV
regression. In Donald W.K. Andrews and James H. Stock, eds. Identification and
Inference for Econometric Models: Essays in Honor of Thomas Rothenberg.
Cambridge: Cambridge University Press.
Summers, Lawrence H., 1986, Does the stock market rationally reflect fundamental
values? Journal of Finance 41, 591-601.
Sunstein, Cass R., 2006, Infotopia: How Many Minds Produce Knowledge? Oxford
University Press.
Tetlock, Paul C., 2004, How efficient are information markets? Evidence from an online
exchange, Yale University Working Paper.
Tetlock, Paul C., and Robert W. Hahn, 2007, Optimal liquidity provision for decision
makers, Yale University Working Paper.
Tetlock, Paul C., 2007, All the news that’s fit to reprint: Do investors react to stale
information? Yale University Working Paper.
Thaler, Richard, and William Ziemba, 1988, Anomalies: parimutuel betting markets:
racetracks and lotteries, Journal of Economic Perspectives 2, 161-174.
Tversky, Amos, and Daniel Kahneman, 1992, Advances in prospect theory: cumulative
representation of uncertainty, Journal of Risk and Uncertainty 5, 297-323.
Wald, Abraham, 1943, Tests of statistical hypotheses concerning several parameters
when the number of observations is large, Transactions of the American
Mathematical Society, 54, 426-482.
White, Halbert, 1980, A heteroskedasticity-consistent covariance matrix estimator and a
direct test for hetereoskedasticity, Econometrica 48, 817-838.
Wolfers, Justin, and Eric Zitzewitz, 2004, Prediction markets, Journal of Economic
Perspectives 18, 107-126.
Wu, George, and Richard Gonzalez, 1996, Curvature of the probability weighting
function, Management Science 42, 1676-1690.
Wurgler, Jeffrey, and Ekaterina Zhuravskaya, 2002, Does arbitrage flatten demand curves
for stocks? Journal of Business 75, 583-608.
Zitzewitz, Eric, 2006, Price discovery among the punters: using new financial betting
markets to predict intraday volatility, Dartmouth University Working Paper.
37
Table I: Summary of Liquidity Statistics for Prediction Market Securities
All securities are based on one-day binary outcome events: either sports games or levels
of financial indexes. At 30-minute intervals, a web crawler records statistics for all
securities in which active trading occurs from March 17, 2003 to October 23, 2006. Table
I summarizes three measures of securities market liquidity: quoted bid-ask spreads,
market depth, and cumulative trading volume—see text for definitions. The table reports
the pair-wise correlations between the logarithms of quoted spreads (LnSpread), market
depth (LnDepth), and trading volume (LnVolm) for all securities, sports securities and
financial securities. The means for each liquidity measure are equal-weighted (EW)
across all order book observations.
Securities Included All Sports Financial
Correlations LnSpread-LnDepth -0.434 -0.334 -0.208
LnSpread-LnVolm -0.012 -0.101 -0.205
LnDepth-LnVolm 0.124 0.177 0.414
Means EW Spread 2.51 2.08 4.37
EW Depth 537 605 238
EW Volume 455 405 674
# of Observations 247613 201492 46121
# of Securities 46670 35089 11581
38
Table II: Returns to Expiration for Sports and Financial Securities
This table reports the results from three ordinary least squares (OLS) regressions of
securities’ returns to expiration on five dummy variables, representing five equally-
spaced pricing intervals: (0,20), [20,40), [40,60), [60,80), and [80,100) points.. The three
regressions include observations of the returns and prices of securities based on one-day
sports, financial, and all events recorded at 30-minute intervals in which there is active
trading. I compute returns to expiration as the payoff at expiration, 0 or 100 points, minus
the bid-ask midpoint divided by 100 points. The small probabilities row displays the
magnitude and significance of the average coefficient on Price1 and Price2. The large
probabilities row displays the magnitude and significance of the average coefficient on
Price3, Price4, and Price5. The large minus small row reports the magnitude and
significance of the difference in these two averages. Standard errors clustered for
securities that expire on the same calendar day appear in parentheses (Froot (1989)). The
symbols *, **, and *** denote significance at the 10%, 5%, and 1%, levels, respectively.
Sports Financial All
0 < Price < 20 -1.65** -0.30 -0.55
(0.67) (0.63) (0.53)
20 ≤ Price < 40 -0.64 -0.23 -0.46
(1.25) (1.12) (0.87)
40 ≤ Price < 60 -0.65 2.47* -0.48
(0.53) (1.41) (0.50)
60 ≤ Price < 80 0.82 2.81** 1.09
(0.88) (1.42) (0.78)
80 ≤ Price < 100 1.73** 1.34 1.60***
(0.81) (0.82) (0.59)
Small Probabilities -1.15* -0.27 -0.50
(0.75) (0.79) (0.59)
Large Probabilities 0.63 2.21** 0.74*
(0.52) (1.02) (0.44)
Large – Small 1.78** 2.47** 1.24*
(0.83) (1.03) (0.68)
R2 0.0003 0.0016 0.0003
Expiration Days 1110 707 1122
Observations 201492 46121 247613
39
Table III: Returns to Expiration for Sports and Financial Securities Sorted by Bid-Ask Spread
This table reports the results from eight (2x4) ordinary least squares (OLS) regressions of securities’ returns to expiration on five
dummy variables (Price1 to Price5), representing five equally-spaced pricing intervals: (0,20), [20,40), [40,60), [60,80), and [80,100).
The eight regressions include two sets of regressions for securities based on one-day binary outcome sports and financial events. Each
set includes regressions for groups of securities sorted into four quantiles by their bid-ask spreads (S). The small, large, and small
minus large probabilities rows display the magnitude and significance of the average coefficients on Price1 and Price2, Price3, Price4,
and Price5, and their difference, respectively. Standard errors clustered for securities that expire on the same calendar day appear in
parentheses (Froot (1989)). The symbols *, **, and *** denote significance at the 10%, 5%, and 1%, levels, respectively.
Sports Securities Financial Securities
Bid-Ask Spread (S) S ≤ 1 1 < S ≤ 3 3 < S ≤ 5 S > 5 S ≤ 1 1 < S ≤ 3 3 < S ≤ 5 S > 5
0 < Price < 20 -3.42*** -1.15 -1.35 -0.91 -0.09 -0.61 -0.19 -0.31
(1.26) (1.00) (1.01) (1.40) (1.34) (0.76) (0.73) (0.77)
20 ≤ Price < 40 -2.22 -0.58 1.06 -2.64 -5.05 2.05 0.71 -1.15
(3.67) (1.42) (1.52) (1.62) (4.64) (1.92) (1.50) (1.14)
40 ≤ Price < 60 -1.06 -0.45 -0.84 -0.89 5.87 2.26 -0.47 3.98***
(1.16) (0.58) (0.55) (0.88) (5.43) (2.35) (1.81) (1.50)
60 ≤ Price < 80 0.92 0.88 0.19 1.94 2.41 5.15** 0.97 3.33**
(1.77) (0.95) (1.05) (1.38) (4.93) (2.17) (1.87) (1.53)
80 ≤ Price < 100 2.79 1.39 1.45 2.28** 2.33** 1.23 1.41 1.24
(1.87) (1.07) (0.96) (0.98) (0.92) (1.21) (0.88) (1.05)
Small Probabilities -2.82 -0.87 -0.15 -1.77 -2.57 0.72 0.26 -0.73
(1.92) (0.93) (0.96) (1.12) 2.47 (1.16) (1.00) (0.83)
Large Probabilities 0.88 0.61 0.27 1.11 3.54 2.88** 0.64 2.85***
(1.06) (0.59) (0.57) (0.68) (2.42) (1.41) (1.23) (1.10)
Large – Small 3.70* 1.47 0.41 2.89** 6.11* 2.16 0.37 3.58***
(2.20) (1.03) (1.03) (1.24) (3.45) (1.51) (1.22) (1.18)
R2 0.0008 0.0002 0.0003 0.0011 0.0096 0.0038 0.0004 0.0033
Expiration Days 634 1099 1100 998 351 616 669 696
Observations 30818 112839 43380 14455 839 6218 15811 23253
40
Table IV: The Effect of Liquidity on the Convergence of Prices to Outcomes
This table reports the results from six two-stage least squares regressions of the reduction in squared percentage returns to expiration
on liquidity and several control variables. I compute the reduction in squared returns over the 30-minute period following liquidity and
control variable measurement. The key independent variables are the logarithms of bid-ask spreads (LnSpread), market depth
(LnDepth), and trading volume (LnVolm). I use two dummy variables to instrument for changes in liquidity using TradeSports market
design changes on September 1, 2004 and November 9, 2004. Each regression uses observations from March 1, 2004 to February 9,
2004. I include a control variable for an event’s expected information release through expiration, ExpInfo, which is (100 – Price) *
Price. All regressions also include controls for six event time dummies, an event time trend, and a date trend. The Anderson, Cragg-
Donaldson, and Hansen test statistics for underidentification, weak identification, and overidentification appear below, along with
appropriate p-values and critical values. Standard errors clustered for securities that expire on the same calendar day appear in
parentheses (Froot (1989)). The symbols *, **, and *** denote significance at the 10%, 5%, and 1%, levels, respectively.
MSPE Reduction in Sports MSPE Reduction in Financial
LnSpread 310*** 978**
(90) (445)
LnDepth -284*** -465*
(85) (242)
LnVolm -189*** -188*
(47) (108)
ExpInfo 0.23*** 0.25*** 0.13*** 0.13* 0.23*** 0.24***
(0.04) (0.04) (0.04) (0.07) (0.04) (0.04)
Anderson LR 711.8*** 541.3*** 405.4*** 42.8*** 125.5*** 151.3***
Anderson p-value 0.000 0.000 0.000 0.000 0.000 0.000
Cragg-Donaldson F 364.4*** 275.5*** 205.3*** 21.5*** 63.7*** 77.1***
F for 10% IV Size 19.9 19.9 19.9 19.9 19.9 19.9
Hansen J 0.15 3.07* 0.02 0.65 1.16 2.52
Hanson p-value 0.694 0.080 0.884 0.420 0.282 0.112
R2 0.231 0.226 0.225 0.010 0.075 0.090
Expiration Days 232 232 232 163 163 163
Observations 14644 14644 14644 3294 3294 3294
41
Table V: Market Order Performance after Informative and Uninformative Periods
Table V summarizes the average daily returns to expiration of market orders that trigger
trades on the TradeSports exchange. The columns report these statistics for the trades
during ―Info‖ periods, during ―No Info‖ periods, and all trades. I form these three groups
separately within sports and financial securities. The ―Info‖ trades are those occurring
during the event and after three trades occurring within 60 seconds of each other. The
―No Info‖ trades are those occurring before the event and after three trades that did not
occur within 10 minutes of each other. On each day, I aggregate the properties of the
trades in all groups from the perspective of the market order trader. The value-weighted
(VW) statistics weight trades by the number of securities traded, whereas equal-weighted
(EW) statistics weight trades equally. The ―EW Return w/ Lag‖ row computes the
hypothetical equal-weighted return to expiration for a trader taking the same position in
the current trade as the previous market order trader. I define the VW buy-sell imbalance
as the quantity of buyer-initiated trades minus the quantity of seller-initiated trades
divided by the total quantity of trades. Standard errors in parentheses are robust to
heteroskedasticity and autocorrelation up to five lags (Newey and West (1987)).
Sports Financial
Info No Info All Info No Info All
VW Return 1.20*** -0.77** 0.38*** 1.41*** -1.12** 0.35***
(0.28) (0.34) (0.12) (0.30) (0.56) (0.10)
EW Return 0.48** -0.54*** 0.07 0.94*** -0.98*** 0.29***
(0.22) (0.18) (0.08) (0.17) (0.34) (0.06)
EW Return w/ Lag 0.91*** -0.16 0.26*** 0.53*** -0.12 0.34***
(0.21) (0.18) (0.09) (0.18) (0.35) (0.06)
VW Buy-Sell Imbal 0.005 0.217*** 0.067*** 0.013 0.062*** -0.001
(0.008) (0.008) (0.004) (0.009) (0.015) (0.003)
# Trades 308 208 1485 303 48 1429
Days 1518 1540 1546 1080 1051 1089
42
Table VI: Market Order Imbalances and Returns Sorted by Pricing Quantile
Table VI summarizes the average daily returns to expiration and buy-sell imbalances of
market orders that trigger trades on the TradeSports exchange. I sort both sports (Panel
A) and financial (Panel B) securities into five equally-spaced pricing quantiles—(0,20),
[20,40), [40,60), [60,80), and [80,100)—using the securities price from the previous
transaction. For each of these ten (2x5) groups of securities, I aggregate each day’s buy-
sell imbalance and returns for all trades taking place during the scheduled event time. I
compute the properties of the trades in the ten groups from the perspective of the market
order trader. The value-weighted (VW) statistics weight trades by the number of
securities traded. Table VI reports these daily averages of trading statistics for the ten
groups of securities. I define the VW buy-sell imbalance as the quantity of buyer-initiated
trades minus the quantity of seller-initiated trades divided by the total quantity of trades.
Standard errors in parentheses are robust to heteroskedasticity (White (1980)).
Panel A: Sports Securities
0 < P < 20 20 ≤ P < 40 40 ≤ P < 60 60 ≤ P < 80 80 ≤ P < 100
VW Buy-Sell Imbal -0.167*** -0.112*** 0.059*** 0.127*** 0.166***
(0.009) (0.008) (0.005) (0.007) (0.008)
VW Return -0.04 1.44*** 0.64*** 0.89*** 0.48*
(0.31) (0.35) (0.23) (0.29) (0.26)
# Trades 153 178 412 242 160
# Days 1511 1531 1543 1539 1515
Panel B: Financial Securities
0 < P < 20 20 ≤ P < 40 40 ≤ P < 60 60 ≤ P < 80 80 ≤ P < 100
VW Buy-Sell Imbal -0.016* -0.081*** 0.000 0.074*** 0.054***
(0.009) (0.006) (0.006) (0.008) (0.009)
VW Return 0.03 0.21 0.74*** 0.76*** 0.40
(0.25) (0.28) (0.25) (0.27) (0.31)
# of Trades 298 302 298 228 198
# of Days 1082 1081 1080 1075 1059
43
Table VII: Unexecuted Limit Orders Sorted by Pricing Quantile
Table VII reports average daily statistics from the TradeSports order book recorded at 30-
minute intervals. I sort both sports (Panel A) and financial (Panel B) securities into five
equally-spaced pricing quantiles—(0,20), [20,40), [40,60), [60,80), and [80,100)—using
the securities price from the previous transaction. For each of these ten (2x5) groups of
securities, I aggregate each day’s buy-sell imbalance and order book liquidity across all
data recording periods. The value-weighted (VW) statistics weight observations by order
book depth, whereas the equal-weighted (EW) statistics weight observations equally. See
text for details on buy-sell imbalance, depth, volume, and spread definitions. Standard
errors in parentheses are robust to heteroskedasticity (White (1980)).
Panel A: Sports Securities
0 < P < 20 20 ≤ P < 40 40 ≤ P < 60 60 ≤ P < 80 80 ≤ P < 100
VW Buy-Sell Imbal 0.045* 0.085*** 0.012*** 0.034*** 0.117***
(0.026) (0.012) (0.004) (0.007) (0.014)
EW Buy-Sell Imbal 0.051** 0.087*** 0.019*** 0.047*** 0.138***
(0.026) (0.011) (0.004) (0.006) (0.014)
EW Depth 394*** 476*** 641*** 616*** 475***
(13) (16) (9) (10) (10)
EW Volume 3393*** 1683*** 422*** 640*** 1629***
(168) (95) (17) (36) (82)
VW Bid-Ask Spreads 2.67*** 2.56*** 1.93*** 2.00*** 2.48***
(0.07) (0.05) (0.02) (0.03) (0.04)
EW Bid-Ask Spreads 2.75*** 2.74*** 2.13*** 2.20*** 2.62***
(0.07) (0.05) (0.02) (0.03) (0.04)
# of Days 490 784 1089 1001 745
Panel B: Financial Securities
0 < P < 20 20 ≤ P < 40 40 ≤ P < 60 60 ≤ P < 80 80 ≤ P < 100
VW Buy-Sell Imbal 0.030** 0.117*** 0.134*** 0.158*** 0.278***
(0.016) (0.010) (0.008) (0.011) (0.018)
EW Buy-Sell Imbal 0.000 0.130*** 0.161*** 0.185*** 0.303***
(0.017) (0.010) (0.009) (0.011) (0.017)
EW Past Depth 279*** 278*** 263*** 250*** 238***
(7) (7) (6) (6) (6)
EW Past Volume 1009*** 990*** 1103*** 1106*** 1335***
(55) (37) (50) (48) (65)
VW Bid-Ask Spreads 3.95*** 4.32*** 4.29*** 4.29*** 3.99***
(0.05) (0.05) (0.05) (0.06) (0.06)
EW Bid-Ask Spreads 4.08*** 4.49*** 4.53*** 4.51*** 4.11***
(0.05) (0.05) (0.05) (0.05) (0.06)
# of Days 516 535 581 585 522
44
Figure 1: The Favorite-Longshot Bias in Sports and Financial Securities
This figure depicts the estimated returns to expiration of securities based on one-day
events for five equally-spaced pricing quantiles: (0,20), [20,40), [40,60), [60,80), and
[80,100). The three series in the figure depict the returns of securities based on sports
events, financial events, and all events combined. Thus, the figure plots the three sets of
coefficient estimates on the five pricing category coefficients shown in the three columns
in Table II (see table for construction).
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
3
1 2 3 4 5
Ret
urn
s to
Ex
pir
atio
n
Pricing Quantile
Sports
Financial
All
45
Figure 2: The Effect of Liquidity on Returns to Expiration
This figure depicts the estimated favorite-longshot biases in returns to expiration of sports
and financial securities with differing degrees of liquidity as measured by three proxies—
bid-ask spreads, market depth, and trading volume. I measure favorite-longshot biases for
securities in each of the four bid-ask spread quantiles, where the bias is the expected
returns on favorites minus the expected returns on longshots—see Table III and the text
for details. I perform analogous tests for securities within each market depth and trading
volume quantile.
-1.00
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
1 2 3 4
Fav
ori
te-L
ongsh
ot
Bia
s in
Ret
urn
s to
Ex
pir
atio
n
Liquidity Quantile
Sports Spreads
Financial Spreads
Sports Depth
Financial Depth
Sports Volume
Financial Volume
46
Figure 3: Market Order Performance after Informative and Uninformative Periods
This figure summarizes the average daily returns to expiration for market orders that
trigger trades on the TradeSports exchange. Figure 3 reports these statistics for the trades
during ―Info‖ periods, during ―No Info‖ periods, and all trades. I form these three groups
separately within sports and financial securities. The ―Info‖ trades are those occurring
during the event and after three trades occurring within 60 seconds of each other. The
―No Info‖ trades are those occurring before the event and after three trades that did not
occur within 10 minutes of each other. On each day, I aggregate the properties of the
trades in all groups from the perspective of the market order trader. The value-weighted
(VW) return series weights trades by the number of securities traded, whereas equal-
weighted (EW) return series weights trades equally. The ―EW Return w/ Lag‖ data series
computes the hypothetical equal-weighted return to expiration for a trader taking the
same position in the current trade as the previous market order trader.
-1.50
-1.00
-0.50
0.00
0.50
1.00
1.50
Sports
Info
Sports
No Info
Sports
All
Financial
Info
Financial
No Info
Financial
All
Ret
urn
s to
Ex
pir
atio
n i
n P
erce
nt
Group of Securities
VW Return
EW Return
EW Return w/ Lag
47
Figure 4: Buy-Sell Order Imbalances Sorted by Pricing Quantiles
This figure depicts value-weighted buy-sell imbalances of four different order types
sorted by the prices of the securities. For limit orders, observations come from the
TradeSports order book recorded at 30-minute intervals, and value weights are based on
market depth. For market orders, observations come from the TradeSports transaction
archive, and value weights are based on the quantity of securities traded. I sort both
sports and financial securities into five equally-spaced pricing quantiles—(0,20), [20,40),
[40,60), [60,80), and [80,100)—using the securities price from the previous order book
recording or previous transaction. For each of these ten (2x5) groups of securities, I
aggregate each day’s market or limit order buy-sell imbalance. See Table VI and Table
VII for further details.
-20.0%
-15.0%
-10.0%
-5.0%
0.0%
5.0%
10.0%
15.0%
20.0%
25.0%
30.0%
35.0%
1 2 3 4 5
Buy-S
ell
Ord
er I
mbal
ance
in P
erce
nt
Pricing Quantile
Sports Market
Financial Market
Sports Limit
Financial Limit